Kevin Wolf <[email protected]> writes: > During shutdown, blockdev_close_all_bdrv_states() drops any block node > references that are still owned by the monitor (i.e. the user). However, > in doing so, it forgot to also remove the node from monitor_bdrv_states > (which qmp_blockdev_del() correctly does), which means that later calls > of bdrv_first()/bdrv_next() will still return the (now stale) pointer to > the node. > > Usually there is no such call after this point, but in some cases it can > happen. In the reported case, there was an ongoing migration, and the > migration thread wasn't shut down yet: migration_shutdown() called by > qemu_cleanup() doesn't actually wait for the migration to be shut down, > but may just move it to MIGRATION_STATUS_CANCELLING. The next time > migration_iteration_finish() runs, it sees the status and tries to > re-activate all block devices that migration may have previously > inactivated. This is where bdrv_first()/bdrv_next() get called and the > access to the already freed node happens. > > It is debatable if migration_shutdown() should really return before > migration has settled, but leaving a dangling pointer in the list of > monitor-owned block nodes is clearly a bug either way and fixing it > solves the immediate problem, so fix it. > > Cc: [email protected] > Reported-by: Thomas Huth <[email protected]> > Signed-off-by: Kevin Wolf <[email protected]>
Reviewed-by: Fabiano Rosas <[email protected]>
